Abstract As the primary electrochemical energy storage device in electric vehicles, the health status of lithium-ion batteries plays a crucial role in determining battery longevity and operational safety. To facilitate early detection of potential anomalies in battery packs, we introduce a multi-step anomaly warning methodology grounded in multi-dimensional feature fusion and the Echo State Enhanced Network with Kolmogorov-Arnold Framework(ESANet). Initially, real-world electric vehicle data are preprocessed, and sensor-based features, along with time-domain and frequency-domain features of battery voltage, are fused. To refine the effectiveness of these feature combinations, the maximum relevance minimum redundancy(mRMR) technique is leveraged for meticulous feature selection. To enhance multi-step voltage prediction, ESANet integrates an echo state network(ESN) to model temporal dependencies and the Kolmogorov-Arnold Framework(KAN) to capture nonlinear dynamics. Furthermore, an anomaly detection system is devised, employing extreme value entropy and predefined thresholds based on the forecasted maximum and minimum voltages. Experimental results validate the performance of the method across normal and faulty vehicles, achieving an impressive average anomaly prediction accuracy of up to 99.87\% within a one-minute horizon. The proposed method offers an effective solution for anomaly detection and fault warning in battery management systems(BMS).
Wu et al. (Fri,) studied this question.